NDArray.ScalarType.bfloat16

Declaration

case bfloat16

Discussion

BFloat16 (Brain Floating Point) uses 8 exponent bits and 7 mantissa bits, matching the exponent range of Float32 but with reduced precision. This type is widely supported across most operations including matrix multiplication, activation functions, and quantization.

See Also

Defining floating-point types

• NDArray.ScalarType.float16
• NDArray.ScalarType.float32
• NDArray.ScalarType.float64